2ClassBADetection/attack_utils.py at main · SafeAndSecureAI/2ClassBADetection · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
import sys
import os
import torch
import numpy as np
import random
import matplotlib.pyplot as plt


def create_pattern(im_size, config):

    if config['PATTERN_TYPE'] == "perturbation":
        pert_size = config['PERTURBATION_SIZE']
        pert_shape = config['PERTURBATION_SHAPE']
        if pert_shape == 'chessboard':
            pert = torch.zeros(im_size)
            for i in range(im_size[1]):
                for j in range(im_size[2]):
                    if (i + j) % 2 == 0:
                        pert[:, i, j] = torch.ones(im_size[0])
            pert *= pert_size
        elif pert_shape == 'static':
            pert = torch.zeros(im_size)
            for i in range(im_size[1]):
                for j in range(im_size[2]):
                    if (i % 2 == 0) and (j % 2 == 0):
                        pert[:, i, j] = torch.ones(im_size[0])
            pert *= pert_size
        elif pert_shape == 'lshape':
            pert = torch.zeros(im_size)
            cx = torch.randint(low=3, high=im_size[-1] - 3, size=(1,))
            cy = torch.randint(low=3, high=im_size[-1] - 3, size=(1,))
            for c in range(im_size[0]):
                pert[c, cx, cy + 1] = pert_size
                pert[c, cx - 1, cy] = pert_size
                pert[c, cx - 2, cy] = pert_size
                pert[c, cx, cy] = pert_size
        elif pert_shape == 'cross':
            pert = torch.zeros(im_size)
            cx = torch.randint(low=3, high=im_size[-1] - 3, size=(1,))
            cy = torch.randint(low=3, high=im_size[-1] - 3, size=(1,))
            for c in range(im_size[0]):
                pert[c, cx, cy - 1] = pert_size
                pert[c, cx, cy + 1] = pert_size
                pert[c, cx - 1, cy] = pert_size
                pert[c, cx + 1, cy] = pert_size
                pert[c, cx, cy] = pert_size
        elif pert_shape == 'X':
            pert = torch.zeros(im_size)
            cx = torch.randint(low=3, high=im_size[-1] - 3, size=(1,))
            cy = torch.randint(low=3, high=im_size[-1] - 3, size=(1,))
            ch = torch.randint(low=0, high=im_size[0], size=(1,))
            pert[ch, cx - 1, cy - 1] = pert_size
            pert[ch, cx - 1, cy + 1] = pert_size
            pert[ch, cx + 1, cy - 1] = pert_size
            pert[ch, cx + 1, cy + 1] = pert_size
            pert[ch, cx, cy] = pert_size
        elif pert_shape == 'pixel':
            pert = torch.zeros(im_size)
            cx = torch.randint(low=3, high=im_size[-1] - 3, size=(1,))
            cy = torch.randint(low=3, high=im_size[-1] - 3, size=(1,))
            ch = torch.randint(low=0, high=im_size[0], size=(1,))
            sgn = torch.randint(low=0, high=2, size=(1,)) * 2 - 1
            pert[ch, cx, cy] += sgn * pert_size * (1 + 0.2 * random.random())
        elif pert_shape == 'square':
            pert = torch.zeros(im_size)
            cx = torch.randint(low=3, high=im_size[-1] - 3, size=(1,))
            cy = torch.randint(low=3, high=im_size[-1] - 3, size=(1,))
            ch = torch.randint(low=0, high=im_size[0], size=(1,))
            pert[ch, cx, cy] = pert_size
            pert[ch, cx, cy + 1] = pert_size
            pert[ch, cx + 1, cy] = pert_size
            pert[ch, cx + 1, cy + 1] = pert_size
        else:
            sys.exit("Perturbation shape is unrecognized!")
        return pert
    elif config['PATTERN_TYPE'] == "patch":
        mask_size = config['MASK_SIZE']
        margin = config['MARGIN']
        patch_type = config['PATCH_TYPE']
        if margin * 2 + mask_size >= im_size[1] or margin * 2 + mask_size >= im_size[2]:
            sys.exit("Decrease margin or mask size!")
        # Pick a random location
        x_candidate = torch.from_numpy(np.concatenate([np.arange(0, margin),
                                                       np.arange(int(im_size[1] - margin - mask_size + 1),
                                                                 int(im_size[1] - mask_size + 1))]))
        y_candidate = torch.from_numpy(np.concatenate([np.arange(0, margin),
                                                       np.arange(int(im_size[2] - margin - mask_size + 1),
                                                                 int(im_size[2] - mask_size + 1))]))
        x = x_candidate[torch.randperm(len(x_candidate))[0]].item()
        y = y_candidate[torch.randperm(len(y_candidate))[0]].item()
        # Create mask and pattern
        mask = torch.zeros(im_size)
        mask[:, x:x + mask_size, y:y + mask_size] = 1
        if patch_type == 'noise':
            patch = torch.randint(0, 255, size=(im_size[0], mask_size, mask_size)) / 255
        elif patch_type == 'uniform':
            color = torch.randint(50, 200, size=(im_size[0], 1, 1)) / 255
            patch = torch.ones((im_size[0], mask_size, mask_size)) * color.repeat(1, mask_size, mask_size)
        pattern = torch.zeros(im_size)
        pattern[:, x:x + mask_size, y:y + mask_size] = patch
        pattern = (pattern, mask)
        return pattern
    else:
        sys.exit("Pattern type is unrecognized!")

    pass


def pattern_save(pattern, config, path):

    if config['PATTERN_TYPE'] == "perturbation":
        pattern = pattern.numpy()
        pattern = np.transpose(pattern, [1, 2, 0])
        if config['DATASET'] in ['cifar10', 'cifar100', 'stl10']:
            plt.imshow(pattern)
        else:
            plt.imshow(pattern[:, :, 0], cmap='gray', vmin=0., vmax=1.)
        plt.savefig(os.path.join(path, 'backdoor_pattern.png'))
    elif config['PATTERN_TYPE'] == "patch":
        pattern = pattern[0].numpy()
        pattern = np.transpose(pattern, [1, 2, 0])
        if config['DATASET'] in ['cifar10', 'cifar100', 'stl10']:
            plt.imshow(pattern)
        else:
            plt.imshow(pattern[:, :, 0], cmap='gray', vmin=0., vmax=1.)
        plt.savefig(os.path.join(path, 'backdoor_pattern.png'))
    else:
        sys.exit("Pattern type is unrecognized!")

    pass


def backdoor_embedding(image, pattern, config):

    if config['PATTERN_TYPE'] == "perturbation":
        image += pattern
        image *= 255
        image = image.round()
        image /= 255
        image = image.clamp(0, 1)
    elif config['PATTERN_TYPE'] == "patch":
        image = image * (1 - pattern[1]) + pattern[0] * pattern[1]
    else:
        sys.exit("Pattern type is unrecognized!")

    return image


def poison(trainset, images, labels, ind, config):

    dataset = config['DATASET']
    image_dtype = trainset.data.dtype
    if dataset in ['cifar10', 'cifar100']:
        images = np.rint(np.transpose(images.numpy() * 255, [0, 2, 3, 1])).astype(image_dtype)
        trainset.data = np.concatenate((trainset.data, images))
        trainset.targets = np.concatenate((trainset.targets, labels))
    if dataset in ['mnist', 'fmnist']:
        images *= 255
        images = images.type(image_dtype)
        images = torch.squeeze(images, dim=1)
        trainset.data = torch.cat((trainset.data, images))
        trainset.targets = torch.cat((trainset.targets, labels))
    if dataset == 'stl10':
        images = np.rint(images.numpy() * 255).astype(image_dtype)
        trainset.data = np.concatenate((trainset.data, images))
        trainset.labels = np.concatenate((trainset.labels, labels))
    if dataset in ['cifar10', 'mnist', 'fmnist']:
        trainset.data = np.delete(trainset.data, ind, axis=0)
        trainset.targets = np.delete(trainset.targets, ind, axis=0)

    return trainset